Get This Tool
Ornith-1.0
Pricing
- Model
- Free
Summary
Most open-source coding agents treat the scaffold — the loop structure, tool calls, retry logic — as a fixed given, so the model learns to work around a mediocre harness instead of improving it. Ornith-1.0 is built on the premise that the scaffold itself should be a training target.
The model family spans 9B-Dense through 397B-MoE, all post-trained on Gemma 4 and Qwen 3.5 with a reinforcement learning framework that jointly optimizes the agent scaffold and the solution rollouts it produces. The vendor states benchmark performance on Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw at the top of the open-source range for each size tier. There is no hosted API — you pull the weights and run inference yourself, which means your infra team owns the deployment stack from day one. The 397B-MoE variant requires hardware that most teams do not have on-hand, so realistic entry for self-hosted production starts at the 31B-Dense tier. Community activity is early-stage: ten commits in the repository, six open issues, no closed pull requests.
Bottom line: Pick Ornith-1.0 if you are running SWE-Bench-style evaluations or repository-level coding research and want weights you can modify freely under MIT — but if you need a managed API or production reliability guarantees before your own infra is ready, there is nothing here to hand that responsibility to.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Community Benchmarks Community
Sign in to submit a benchmarkNo community benchmarks yet. Be the first to share a real-world data point.
Pros
Sign in to edit- Four model sizes (9B-Dense to 397B-MoE) targeting different hardware tiers, so teams can match the checkpoint to available GPU memory without being locked into a single model footprint.
- Scaffold optimization via RL means the agent learns better tool-calling and search patterns during training rather than inheriting a fixed harness — which directly addresses the ceiling most fine-tuned coding models hit on multi-step repository tasks.
- MIT license with no stated regional restrictions, so the weights can be used, modified, and redistributed freely — removing the legal friction that blocks open-source derivatives with more restrictive licenses.
- Post-trained on Gemma 4 and Qwen 3.5 base checkpoints, which means teams already running those base models can reason about the training lineage and diff behavior against the base without starting from scratch.
- Self-hosted by design, so inference costs, data privacy, and network dependencies stay inside your own infrastructure rather than flowing through a third-party API.
Cons
Sign in to edit- There is no API, SDK, or hosted inference endpoint — every deployment starts with raw weight files and your own serving stack. Teams without existing LLM inference infrastructure will spend a sprint on setup before writing a single agent task.
- The 397B-MoE checkpoint requires GPU capacity that most product teams do not operate; the docs provide no quantization or offloading guidance in the scraped content, so teams with consumer-grade hardware are effectively limited to the 9B or 31B-Dense tiers.
- The repository shows ten commits and six open issues with zero closed pull requests, indicating early-stage maintenance maturity. Teams that hit a bug in the training loop or scaffold tooling have no established patch cadence to rely on — they fix it themselves or wait.
- There is no observability, logging, or tracing layer included. Teams running production coding agents who need to audit what the scaffold decided at each step have to instrument that entirely from scratch, which is a non-trivial engineering surface.
- Any team that needs a usage API, rate-limit controls, or multi-tenant access management will abandon Ornith-1.0 for a hosted alternative — the architecture makes no provision for those concerns, and adding them requires building a full serving layer on top of the weights.
Community Reviews
Sign in to write a reviewNo reviews yet. Be the first to share your experience.
About
- API Available
- No
- Self-Hosted
- Yes
- Last Updated
- 2026-06-30T13:38:47.404Z
Best For
Who it's for
- Open-source agentic coding workflows
- Self-improving model research
- Benchmark-competitive coding agents
What it does well
- Agentic coding tasks
- SWE-Bench style repository-level fixes
- Terminal-based coding benchmarks
- NL2Repo code generation
Discussion Community
Sign in to commentNo discussion yet. Sign in to start the conversation.
Compare Ornith-1.0
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeCommunity Notes & Tips Community
Sign in to contributeBe the first to contribute. General notes, observations, gotchas, and tips from people who use this tool day-to-day.
Frequently Asked Questions
- Is Ornith-1.0 free?
- Yes — Ornith-1.0 is fully free to use. There is no paid tier.
- Is Ornith-1.0 open source?
- Yes. Ornith-1.0 is open source.
- Can I self-host Ornith-1.0?
- Yes. Ornith-1.0 supports self-hosting on your own infrastructure.
Hours Saved & ROI Stories Community
Sign in to contributeBe the first to contribute. Concrete time/cost savings, with context. e.g. "Cut my code review backlog from 4h to 45m per week."
Curated lists that include this category
Ornith-1.0 is a family of open-source agentic coding models that the vendor describes as self-improving: rather than fixing the agent scaffold and training only the solution policy, the RL training loop treats the scaffold itself as a learnable parameter. In practice, this means the model jointly discovers which search trajectories and tool-calling patterns produce better code, rather than inheriting a hand-engineered harness. The workflow is weight-based — you pull a checkpoint (9B-Dense, 31B-Dense, 35B-MoE, or 397B-MoE), deploy it against your own inference stack, and wire it into your benchmark runner or coding pipeline. No API key, no cloud dependency.
The differentiating architectural claim is the scaffold optimization loop. Standard fine-tuned coding models learn to complete tasks inside a fixed agent harness; Ornith-1.0’s RL framework generates and scores candidate scaffolds alongside solution rollouts. The docs describe this as producing better search trajectories, which the benchmark numbers on SWE-Bench and Terminal-Bench 2.1 are intended to validate at each model tier against size-appropriate baselines.
Where this fits cleanly: research teams benchmarking open-source coding agents, engineers running NL2Repo or OpenClaw evaluations, and teams that need MIT-licensed weights they can fine-tune or redistribute without restriction. Where it breaks: any team that needs a managed API endpoint, rate-limit guarantees, or usage observability out of the box gets none of those here. The 397B-MoE tier also carries a hardware requirement that puts it out of reach for teams without large GPU clusters, making it a research asset rather than a production deployment target for most organizations.
All model weights are MIT licensed with no stated regional restrictions. The repository is hosted on GitHub with the primary access point being the README and assets folder; there is no package registry entry or pip-installable wrapper documented in the scraped content.